Identification of tightly regulated groups of genes during Drosophila melanogaster embryogenesis

Time-series analysis of whole-genome expression data during Drosophila melanogaster development indicates that up to 86% of its genes change their relative transcript level during embryogenesis. By applying conservative filtering criteria and requiring ‘sharp' transcript changes, we identified 1534 maternal genes, 792 transient zygotic genes, and 1053 genes whose transcript levels increase during embryogenesis. Each of these three categories is dominated by groups of genes where all transcript levels increase and/or decrease at similar times, suggesting a common mode of regulation. For example, 34% of the transiently expressed genes fall into three groups, with increased transcript levels between 2.5–12, 11–20, and 15–20 h of development, respectively. We highlight common and distinctive functional features of these expression groups and identify a coupling between downregulation of transcript levels and targeted protein degradation. By mapping the groups to the protein network, we also predict and experimentally confirm new functional associations.


Introduction
For the purpose of tracking relative transcript levels during development, microarray analysis has proven to be invaluable. The partial transcriptomes of two major model organisms have already been analyzed, the nematode Caenorhabditis elegans (Baugh et al, 2003) during embryogenesis, and an extensive developmental time series in Drosophila melanogaster of approximately 30% of all genes covering the entire lifespan, from the first minutes of development to aging adults . The latter study gave the first insights into global changes of regulation, such as the prominent biphasic expression of many genes in two major stages, either in the embryo and pupa, or in the larva and adult, revealing the molecular similarities in these stages of the lifecycle.
Recently, genome-wide expression in fruitfly was measured at the exon level, providing enough resolution to identify alternative splicing in 40% of predicted genes and to identify at least 15% as developmentally regulated (Stolc et al, 2004). However, as only two time points were considered during early and late embryo development, many transient and tightly regulated processes during embryogenesis would not have been detected.
Here, we perform an extensive analysis of the fly transcriptome, comprising 12 868 genes (FlyBase 4.0 release; Drysdale and Crosby, 2005), during 30 time points, covering the entire 24-h period in which the fertilized egg develops into a larva. This enabled us to identify transition points of sharp changes in transcript levels, as well as groups of genes with similar expression profiles during embryo development. Several common functional features were identified among the proteins encoded by genes with similar temporal coexpression patterns. For example, a significant enrichment of physical interactions implies the presence of entire programs of 'effector' genes involved in a common developmental process. Furthermore, we show on a more global level that tight regulation of transcript levels is often accompanied by targeted protein degradation. As a result, we increase our understanding of several vital developmental pathways, and we suggest new, as yet unknown, members of these pathways, some of which we validate experimentally.

Results
Generation of the developmental time-series data changes that occur during the first half of embryogenesis, overlapping 1-h time points were obtained for the first 6.5 h of development. The stages of all samples were verified and only tightly staged embryo collections were used for RNA isolation and microarray analysis in order to minimize the overlap between measurements (see Figure 1 for the distribution of stages at each time point).
Three independent embryo collections were used for each stage of development. Details of the microarray hybridizations are found in Arbeitman et al (2002). All samples were hybridized together with a common reference sample, which was made from pooled samples for each transcript from all stages of the Drosophila lifecycle. Thus, the expression level of each gene in the sample can be compared relative to its corresponding reference expression level. The microarrays used for this study consist of PCR fragments of one exon of every predicted Drosophila gene . The data were normalized using the intensity-dependent Qspline method (Workman et al, 2002) and subsequently corrected for spatial biases (see Materials and methods for details).

Estimating expression changes of genes during embryogenesis
Two different statistical methods were used for identifying genes that change in expression during embryonic develop-ment. First, the widely used analysis of variance (ANOVA) was applied to identify significant changes in the general expression level. We found significant changes in transcript levels for 86% (Po0.05) of all genes. This compares to C. elegans (Baugh et al, 2003) and an earlier analysis of D. melanogaster (Arbeitman et al, 2002), where 68 and 95% respectively (Po0.05) of the genes were found to change expression during embryogenesis. However, the actual implementation of ANOVA may differ slightly. Second, to explicitly analyze the temporal dependency of expression levels in individual genes, a runs test was used; it suggests temporal changes in transcript levels for 65% of genes during embryogenesis.
These estimates give a global overview of the amount of change occurring during the entire embryonic development; however, they do not pinpoint when transitions in gene expression programs occur. To get a more exact time measure, we searched for changes in expression levels using local convolution methods (Supplementary Figure S1). More specifically, we required four points of low expression and four subsequent points of high expression (or vice versa) even if the amplitude change was relatively low (see Materials and methods). This type of convolution not only requires a sharp increase or decrease of expression, but also that the change in transcript level is consistent over a period of time, thereby reducing the rate of false positives owing to individual outliers. Using this approach, we mapped 6233 'sharp' changes in transcript levels (2808 increases and 3425 decreases) to time points and developmental stages ( Figure 1). As indicated already in the study of Arbeitman et al (2002), several developmental stages show an increased frequency of transcript level changes during embryogenesis, which can now be confirmed genome-wide. The local convolution analysis also revealed that the increase in the transcript level of one group of genes often coincides with the decrease of transcript expression of another group of genes and vice versa, indicating coordinated waves of expression ( Figure 2).
A flurry of expression changes was observed at 2-3 h (embryo stage B5), representing the initiation of zygotic transcription and the parallel decay of some maternal transcripts. This first stage of dramatic change coincides with events just after cellularization-the process during which each nucleus is enclosed to form a cell by invagination of the plasma membrane-for example, the major morphogenetic changes leading to germ layer formation that occur during that time period in the cellularized embryo. It might also reflect the embryo patterning that begins along both the anterior-posterior and dorsal-ventral axis (Wolpert et al, 2002).
A second major period of transcript expression change (both increase and decrease) was observed at roughly 12 h, when most embryos had reached stages 12-14, corresponding to the end of the dorsal closure, the terminal differentiation of many tissues, and to the invagination of the epithelial cells that will become the imaginal discs.
A third period of gene expression change is observed at 16 h (stages 14-16) when a discrete set of transcripts decrease their expression levels, followed by an intense increase of transcript levels of another set of genes (17-19 h). This could possibly be in preparation for the transition to the larval stage. Generally, sharp decreases of mRNAs seem to be more confined to particular time points.
The correlation between times of increase and decrease in transcript levels suggests the existence of coregulated groups of genes that drive major developmental events during embryogenesis.

Classification of gene expression behavior
To be able to group genes whose transcripts follow similar patterns over the full 30 time points, and to correlate this with the periods of rapid expression changes, we used global convolution (see Materials and methods and Supplementary Figure S1) to assign genes to general expression classes. These are characterized by distinct plateaus of low and high expression during embryogenesis: class I (maternal) genes encoding transcripts that start with a high relative transcript level, which subsequently decreases; class II (transient) genes whose transcripts levels first increase and later decrease, and thus do not seem to be maternally deposited and, Finally, class III (activated) genes encoding transcripts for which we only observe an increase in expression.
Class III gene transcripts are most likely not present at high levels during the entire Drosophila life cycle, as most of the corresponding genes return to low levels at various time points beyond embryogenesis (Arbeitman et al, 2002). Although the transcripts from the transiently expressed genes (class II) may be present in later stages in the larva, pupa, or the adult fly, we expect only a few genes with multiple expression peaks during embryogenesis (Arbeitman et al, 2002), also indicated by the runs test above (data not shown).
Using global convolution (see Materials and methods), we found strong and significant expression correlation coefficients (r40.8, Po10 À4 , t-test) between transcript levels and global convolution profiles for 26% (3379) of the transcripts present on the array. Of these, we classified 1534 as class I, 792 as class II, and 1053 as class III. Many genes do not fit to any of these three classes, for instance genes that are constitutively transcribed (or not transcribed at all) during embryogenesis. Furthermore, the requirements for assignment were rather stringent; the entire expression profile must fit the categorization to a high degree (30 time points), whereas in the local Figure 2 Major classes of transcript levels, as determined by global convolution. Arcs represent the dominant subgroups within class I, II, and III transcripts. For instance, class I is dominated by two main subgroups I:a and I:b, represented by the pink and red arcs, respectively. Time is in hours, and yellow rectangles signify measurement points. The time of increase and decrease of the transcript groups coincides with those derived by local convolution (Supplementary Figure S1). Note the interplay between groups as a decrease of one transcript group is followed by an increase of another.
convolution, only eight time points are considered (corresponding to Po10 À4 , t-test). This leads to a low rate of false positives at the cost of sensitivity. For instance, we are likely to miss many genes with very short periods of transcriptional activity.
Within each of the three global classes, groups of genes can now be readily identified with common times of increase and/ or decrease of relative transcript levels ( Figure 1A). More than 62% of the class I (maternal) genes can be classified into two major groups: class I:a and b, whose transcript levels decrease at 3-5 and 12-14 h, respectively. For the class II genes, even though there are 276 possible combinations of time points of increase and decrease of transcript levels (see Materials and methods), as many as 38% of the 792 class II genes fall into only three groups: II:a (2.5-12 h), II:b (11-20 h), and II:c (15-20 h) containing 153, 100, and 50 genes, respectively. Of the genes whose transcript levels increase but are not observed to decrease during embryogenesis (class III), more than 73% can be classified into three main groups; class III:a, b, and c (times of increase at 13-14, 18-20, and 22 h of development).
In order to identify biological principles underlying these different coexpression groups, and also as a general quality control, we studied these groups using various sources of biological information (Table Ia-c and Materials and methods). For example, proteins encoded by genes assigned to each of the three major classes have more interactions with each other than expected from a random group of the same size (Po0.01; Table Ia -c), showing that the gene products are not only coexpressed but also tend to interact physically to perform related cellular functions.

Class I: maternal genes
Class I contains 1534 genes, and thus represents the most populated of the three major expression classes. It is long been known that a large number of transcripts are deposited in the oocyte during gametogenesis. Among other vital functions, they have been shown to be responsible for establishing the major body axes and for the initiation of zygotic transcription (Luschnig et al, 2004). The importance of these genes is reflected by a high proportion of lethal genes (observed: 10%; expected: 5%; Po10 À3 ; w 2 -test) and a higher than average fraction of orthologs in this group shared with Anopheles gambiae (Table Ia), indicating functional conservation. Furthermore, analysis of Flybase GO annotation (Drysdale and Crosby, 2005) revealed that there is a significant overrepresentation of genes involved in 'nuclear organization and biogenesis', 'nuclear mRNA splicing', and 'DNA metabolism' in the class I group. These genes facilitate the organization of the chromosomes and nuclei during the very rapid cell divisions in the precellular blastoderm embryo. Proteins encoded by the class I genes also display a strong physical interconnectivity: 1097 connections within 1534 proteins (expected connections: 517; Po10 À3 ; w 2 -test). Furthermore, they are enriched in transcription factors (Po0.05). A total of 365 out of the 1534 genes change transcript levels at 1.5-3 h (group class I:a; Figure 2). As expected, the functionally characterized genes of the class I:a group are mostly involved in early development and in the cell cycle according to the interactive fly database (Brody, 1999). The class I:b group consists of 593 genes, which encode transcripts with decreased levels of expression by 10-11 h. These genes are annotated in the interactive fly database with functions involved in the cell cycle, chromatin organization, and DNA replication. Class I:b is also enriched in lethal genes (Po0.01; w 2 -test). In developmental terms, the decrease of the transcripts in class I:a coincides with the start of gastrulation, and I:b with the initiation of germ band retraction and dorsal closure.

Class II: transient genes
Despite the stringent requirements of class II genes to have both a sharp increase and decrease in transcript levels, 792 fly Orth Orth: proportion of genes with orthology to A. gambiae. Lethal: proportion of genes annotated as 'Phenotypic class: Lethal'. Transcription factors: proportion of genes annotated as 'transcription factor'. Interaction: proportion of genes coding for proteins with known or predicted protein interactions in STRING (see Materials and methods). Pathway1: proportion of genes (counted once) involved in known pathways. PathwayX: as Pathway1, but genes may be counted any number of times, if it appears in several pathways. GO: annotation for the categories molecular function, biological process, and cellular compartment. Regulated genes during fly embryogenesis SD Hooper et al genes were classified in this class. This is consistent with earlier reports in D. melanogaster (Arbeitman et al, 2002) and also C. elegans (Baugh et al, 2003). As shown in Table Ib, the class II genes are enriched in transcription factors and lethal phenotype genes (hereafter referred to as lethals; see Materials and methods) and not surprisingly, this class shows an overrepresentation of genes with functions involved in development. More specifically, many of these genes are annotated as encoding proteins involved in 'histogenesis', 'organogenesis', 'ectoderm development', 'cell differentiation', and 'cell fate commitment' (Table Ib; GO analysis). The class II group is also enriched in genes that encode well-characterized transcription factors (Po0.05, t 2 -test).
Of the class II genes, 303 (38%) are found in only three groups; a, b, and c, each defined by specific times of increase and decrease of transcript levels ( Figure 2). In addition to the common features above, the groups also differ from each other. For example, more genes from class II:a (with a plateau of increased transcript levels starting at 3-6 h and decreasing at 12-13 h) have been functionally characterized than average for the genome (Po0.01, t 2 -test). Conversely, genes in class II:b (13-14 to 19-21 h) and II:c (17-18 to 21 h) are poorly characterized. Class II:a corresponds roughly to the time of cellularization and gastrulation, up to the point of germ band retraction. Class II:b and II:c range between germ band retraction and late embryonic stage.
The class II:a group (153 genes, of which 109 are annotated) contains 11 genes (of 29 on the array) that are a part of the Notch pathway (see below). Some other Notch members are found in class I, as many are maternally inherited. The drop in their transcript levels do however tend to coincide with II:a. The remaining 142 genes are often annotated as being implicated in 'neuroblast cell fate determination' suggesting that this group is highly enriched for both the regulators   (members of the Notch pathway) and their downstream effector molecules (genes involved in neurogenesis). Genes involved in dorso-ventral patterning are also significantly overrepresented in the class II groups (Po0.01) and in particular in the class II:a category (Po0.05). Genes from the class II:a group are expressed in the procephalic ectoderm, ventral ectoderm, sensory complex, and central brain neurons (in situ data, see Materials and methods). Again, this strongly suggests a role for class II:a genes in nervous tissue and brain development.
Despite the low proportion of functionally characterized genes in the class II:b group (51 of 100 genes), there is an enrichment in oxidoreductase and peroxidase functions (GO classification). Moreover, defense and immune response are common functional classifications in this group. For instance, eight (of 20 known) members of the Osiris cluster (Dorer et al, 2003) are present in class II:b. This cluster is part of the largest region of synteny between D. melanogaster and A. gambiae, and encodes one of the largest gene families in fruitfly (Zdobnov et al, 2002). The genes of the Osiris cluster are still poorly characterized, but they are known to be under strong selection pressure both on protein sequence and expression level (Dorer et al, 2003). The transcript levels of Osiris genes 3, 7, 9, 17, 18, and 20 are known to be high during embryogenesis from stages 13 to 16 (Dorer et al, 2003), compatible with our findings.
The class II:c group encompasses 50 genes. Only 19 of these are annotated, and most are involved in cuticle formation. The fly embryo secretes a hard proteinaceous material, which forms a thick protective cuticle surrounding the larvae. Furthermore, class II:b (see above) is linked via in situ data to the dorsal ectoderm, suggesting that also many genes in class II:b may contribute to cuticle formation (Ostrowski et al, 2002). Therefore, groups II:b and II:c may be of interest when designing insecticides or planning experiments that target the cuticle.

Class III: activated genes
Of the 1053 transcripts with sharp increases in expression levels but without a subsequent decrease during embryogenesis (class III), 250 are activated at 11-12 h (III:a), 342 at 16-18 h (III:b), and 184 at 20 h (III:c). These genes are the most species specific of the three categories: we found significantly fewer orthologs to predicted genes from A. gambiae than for the genes represented on the microarray as a whole. Furthermore, known transcription factors are significantly underrepresented in this group (Table Ic), which implies that the mRNA expression of transcriptional regulators is likely to be under tight regulation. III:a starts roughly at the same time as the dorsal closure, and III:b coincides with the late embryonic stage. The initiation of III:c cannot be correlated to any distinct developmental event.

Coordinated regulation of transcripts and protein products
The decrease in transcript levels during embryogenesis in the class I and II genes suggests that it is important to reduce the levels of the respective protein in a temporally controlled manner. Transcriptional regulation alone is not sufficient to ensure a rapid reduction in protein levels, as the protein degradation may take hours. We thus hypothesize that the protein products of most transcriptionally repressed genes are inactivated, for example through targeted degradation controlled by PESTregions. This mechanism has previously been suggested to be responsible for the degradation of maternal proteins in C. elegans (Baugh et al, 2003) as well as for proteins that are periodically expressed during the mitotic cell cycle in several eukaryotes (de Lichtenberg et al, 2005;Jensen et al, 2006).
To test our hypothesis, we used a computational method to systematically predict PEST regions in the D. melanogaster proteome and compared the percentage of PEST-containing proteins encoded by the genome to that of the genes in each expression class. The highest percentages of PEST-containing proteins are encoded by class I (39%) and class II genes (37%). Both groups are significantly enriched in PEST regions compared to the proteome-wide content (31%) with P-values of 10 À11 and 10 À3 , respectively. The difference between class I and II is not statistically significant. In contrast, PEST regions are found in only 21% of the proteins encoded by class III genes whose RNA levels remain high, which is significantly less than expected at Po10 À9 . These observations are consistent with the hypothesis of a coordinated downregulation of a gene's expression in Drosophila at the level of both their RNA and protein products during embryogenesis, as suggested previously for other organisms (Baugh et al, 2003;de Lichtenberg et al, 2005;Jensen et al, 2006).

Mapping coordinately expressed groups of genes to protein interactions
As suggested above, transcripts in the individual expression groups are tightly coexpressed at the same stages of development and are enriched in particular groups of functional (GO) categories. Consequently, one would expect that the protein products preferentially interact with each other to perform common functions. To test this hypothesis, we analyzed the integrated D. melanogaster protein interaction network from the STRING database, which contains large-scale interaction data from fruitfly yeast two-hybrid screens (Giot et al, 2003), small-scale interactions stored in dedicated databases and extracted from the literature and inferred interactions from different species, all embedded into a unified scoring scheme (von Mering et al, 2005). Indeed, proteins encoded by genes in all groups except class I:a and class III:a have more interactions between themselves than with proteins outside that group (Po0.01). Certain classes of proteins and pathways contribute strongly to this enrichment, including proteins involved in energy production, transcription factors (such as bicoid, FBgn0000166), or members of the Notch pathway ( Figure 3). The latter also indicates that functions are not only performed within a particular coexpression group but also that some larger developmental pathways require a concerted action of genes with different expression profiles.

Exploration of coexpressed genes involved in common pathways: Notch as an example
The Notch signaling pathway regulates cell fate determination, and consists of 61 proneural and neurogenic genes (Brody, 1999). Of these 61 genes, 52 were assayed on the array and 20 fit significantly to the category of class II genes. Thus, 432% of the Notch pathway genes have highly coordinated and transient gene expression, as compared to an expected 6.2% (based on 792 of 12 868 transcripts in class II; Po10 À4 , t-test). This reflects the transient role of this pathway in cell fate specification. Roughly half of these genes are found in II:a, along with Notch itself. Also, we find four pathway members (e.g. deltex: FBgn0000524) in the maternal class I:b, whose transcripts decrease at the same time as class II:a.
Some members of the Notch pathway in class II include big brain (FBgn0000180; an ion channel concentrated at apical adherence junctions), neuralized (FBgn0002561; a gene involved in ubiquitination), and a key transcription factor acting in parallel to the proneural genes, soxneuro (FBgn0029123; Overton et al, 2002).
Notch itself (FBgn0004647) and many of its pathway members are highly expressed up to the 12th hour of development (stages 12-14 in this developmental time series), but then have an abrupt decrease in expression, dropping to very low transcript levels after this point. This sharp decrease of Notch and some other transcripts suggests that the pathway is not needed anymore and should be suppressed, indicating that the major specification events have been completed. This is further indicated by the presence of a PEST sequence in Notch, and the fact that the prolonged transcript expression of Notch gives rise to disease (Joutel and Tournier-Lasserve, 1998).

Coexpressed classes of genes show coordinated expression in vivo
As outlined above, the tight temporal patterns of expression observed by 430% of the Notch family members suggest a coordinated regulation in expression. This observation makes a number of predictions about the expression patterns of the other B100 class II:b genes, the majority of which are poorly characterized. Firstly, class II genes should have transient Figure 3 Major component of a literature-derived protein interaction subnetwork, obtained from the String database at a reliability score of at least 0.3 (von Mering et al, 2005). It reveals that several well-known interacting proteins also show similar expression profiles. Examples are the highlighted vacuolar ATPases, the proteasome, or interactors of Peter Pan. Notch appears as the central node of the network and contributes to the high interconnectivity of the (transient) class II:a group. Some genes with expression profiles very similar to Notch (labeled by red arrows) are currently only loosely associated with the pathway (see text), but might share more functionality with Notch than previously thought. Notch is labeled for reference purposes. Note that unassigned in the legend means that the respective genes belong to the class but not to any of the major subclasses. To explore this complex network in full detail, see the interactive figure and data files in Supplementary information.
expression, initiating at stages 6-7 and decreasing by stage 13. Second, as these genes have coordinated expression with Notch family members, they are highly likely to be colocalized in the same cells. To test these hypotheses, we selected five candidates from the II:a group, for costaining with Delta (FBgn0000463; the ligand of the Notch receptor) by double in situ hybridization. The candidates were selected based on their strength of correlation to the II:a profile, their fold change, and their gene annotation.
Four of the five genes gave specific patterns of expression. We were not able to obtain an in situ hybridization probe for the fifth gene (CG1316). For all four genes tested, their expression initiates early in development (Bstage 7), and is dramatically reduced or absent by stage 13 (Figure 4). There is residual expression in small groups of cells at stage 13 for three genes, for example in the brain (worniu and CG13333), in the foregut (pdm2), and segmentally repeated groups of cells in the ectoderm (CG13333). No expression was observed for CG4440 at stage 13, indicating that this gene is no longer transcribed. Therefore, the temporal window and transient nature of expression of all four genes map to the prediction for class II:a genes.
We next examined the spatial colocalization of these four genes with the Notch pathway, using Delta as a marker. Delta is a membrane-bound ligand for Notch, and is therefore expressed on the 'Notch signal-sending' cell. While Notch itself is a membrane bound receptor on the 'Notch-receiving cell'. As Delta is tethered to the membrane, in contrast to other signaling pathways, the 'Notch signal-sending' cells and 'Notch-receiving' cells are usually adjacent to each other or within the same field of cells. A colocalization of tissue expression in the Deltaexpressing cell or the neighboring Notch-expressing cell would suggest that the transcript may play a role either in the Notch-Delta pathway or in a tightly coordinated parallel pathway. The transcription factors pdm2 and worniu are essential for brain and neural development, but are not known to be linked to either Notch or Delta specifically. Figure 5 shows in situ hybridization of stage 11-12 embryos, which have peak expression for class II:a genes expression. As the Notch pathway is highly active during this time of embryogenesis, Delta has a very broad expression pattern making it problematic to discern specific colocalization. Given this potential difficulty, we could observe colocalization of worniu and pdm2 in the ventral nerve cord and brain. These genes are colocalized in a subset of neuroblasts indicating specific colocalization in these cells at this stage of development. CG13333, a gene of unknown function, has a broad expression in a number of tissues including the developing foregut, hindgut, and trachea. Again, we observed specific colocalization with Delta in the brain. Interestingly, both CG13333 and the second uncharacterized gene CG4440 are expressed in ectodermal strips which are directly adjacent to the Delta-expressing ectodermal strips. This neighboring expression may represent CG13333 and CG4440 expression in Notch-receiving cells, rather than the Delta-sending cells, which would implicate these genes in Notch-regulated processes in development.

Discussion
Our approach focused on sharp changes in transcript levels in order to identify genes that may be subject to tight regulation.

Figure 4
Decline of transcript levels of four class II:a genes as predicted by the array analysis. Transcript levels are high until stage 12, but decline rapidly after stage 13.
The functional signals (i.e. orthology, degree of interaction, etc.) that are shared among expression groups and also the signals that distinguish groups from each other lend support to our analysis strategy, despite inclusion of data of varying quality. Although these signals come at the cost of low sensitivity (groups are likely to have more members than this study can identify), the expression categories and groups are a starting point for exploring different functional features. For instance, we note that the transient class II genes are likely to play vital roles in development, based on the behavior of their transcripts. Overall, we find a strong consistency between the global clustering method, which is conceptually based on time-dependent data, and various sources of purely biological information. This further underlines the advantages of timeseries arrays as opposed to non-temporal studies.
Owing to the low number of genes in some of the groups and to their limited annotation, we were unable to uncover the functionality behind the simultaneous decrease of one expression group and the increase of another (e.g. 12-14 h; class II:a and I:b transcript levels decrease, whereas II:b and III:a increase; Figure 2). However, our analyses clearly reveal the existence of these transitions between groups. The most dramatic of these expression changes appears to occur in the developmental stage 8 (12-14 h; Figure 2). This stage tends to be either the point of increase or decrease of transcript levels for many of the expression groups described above. The expression group class II:a decreases sharply at this point and contains a high number of genes encoding transcription factors, which most likely induce expression of various downstream pathways. Together with class I:b, whose transcript levels also decrease at that time, class II:a is best annotated. This group contains the highest fraction of orthologs and is enriched in lethal genes, suggesting that a coordinated decrease of transcript levels at this stage is essential for embryogenesis. The decrease of class I:b and class II:a transcript levels is followed by a burst of sharp transcript level increases from other groups of genes. There are many possible biological explanations for the distinct transition time between class I:b/II:a and class III:a/II:b. One major developmental event at this time is the end of the cell fate determination phase, coordinated by the Notch pathway. Not only do transcript levels decrease but also the proteins they encode, as we observe an enrichment of PEST motifs in the respective protein sequences indicative of a controlled degradation upon phosphorylation, whereas PEST motifs are underrepresented in proteins from class III. In this study, we have identified roughly 100 genes in class II that may be involved in cell determination, for instance in association with Notch or Delta. For a few of these, we have shown experimentally that this prediction is valid. A more detailed study of in situ patterns may provide insights into many more of these uncharacterized genes.
Taken together, our initial analysis of embryonic gene expression in D. melanogaster not only confirmed a number of known expression patterns but also revealed a surprisingly low number of well-defined expression groups that are sharply Figure 5 Spatial colocalization of worniu, pdm2, CG13333 and CG4440 with Delta in embryos stage 11-12. Columns 1 and 2 show stainings individually and column 3 shows colocalization. CG4440 exhibits an anti-correlation, suggesting colocalization with Notch rather than Delta.
activated and suppressed together. Some of them (class I:b, class II:a) contain many well-characterized genes, whereas others (class II:b, class II:c) reveal common temporal aspects of poorly annotated genes that should help in initiating more targeted functional studies. The coupling of sharp gene suppression and targeted protein degradation suggested here should enable network studies that combine temporal regulation and physical protein interaction.

Microarrays and sample preparation
In this work, we used the same samples as described in detail by Arbeitman et al (2002). A brief summary is as follows. Canton S wildtype embryos were collected at 30 time points over a 24-h time period, with overlapping 1-h time points during the first 6.5 h, followed by hourly sampling. The stages of all samples were verified and only tightly staged embryo collections were used for RNA isolation and microarray analysis. The distribution of stages within each time points is shown in Figure 1. All samples were hybridized to a common reference sample. The reference sample, described by Arbeitman et al (2002), was made from pooled samples from all stages of the Drosophila lifecycle and therefore should represent a median level of expression for all genes in the genome. This serves as a constant denominator to which the relative levels of expression of each gene in the experimental samples can be prepared. The microarrays used for this study consist of PCR fragments of one exon of every predicted Drosophila gene from release one of the genome (for more details, see supplemental material of Li and White, 2003).

Normalization of microarrays
The spot intensities of the two channels (Cy3 and Cy5) on each microarray were individually normalized using the Qspline method (Workman et al, 2002) with a log-normal distribution as target (M¼ln 1000, S¼ln 1000). The channels were further normalized to correct for spatial biases using a Gaussian smoother with sigma¼0.8 (Workman et al, 2002). After adding a regularization background intensity of 100 to the normalized intensities, a log-ratio was calculated for each gene on each spotted array. This value was semi-empirically chosen to make the spread of log-ratios independent of the spot intensities.

Identification of significantly regulated genes
We performed an ANOVA on each gene in order to determine significant changes in expression as has been carried out in earlier studies of time series (Arbeitman et al, 2002;Baugh et al, 2003). The resulting number of regulated genes was high (86% at Po0.05, 70% at Po10 À3 ), although Arbeitman et al (2002) reported even higher numbers (95% at Po0.05 and 86% at Po10 À3 ). The high number of genes with a significant change in expression level could suggest that an ANOVA is not sufficiently specific. We therefore also performed a runs test, which unlike ANOVA takes the temporal ordering of the data points into account and hence fits better with our subsequent analyses. Of 336 known transcription factors (not exclusively involved in embryogenesis), the runs test found 71% of them while suggesting 46% of all genes to be regulated (Po0.05). In comparison, ANOVA found 88% of the transcription factors, but suggested 86% of all genes to be regulated. In this case, ANOVA performs only barely better than random selection.

Local convolution
In order to specify the times of activation and suppression, we convolved array data with vectors of eight integers, for instance x¼[0 0 0 0 1 1 1 1]. We selected those matches with correlation coefficients exceeding 0.9, corresponding to Po10 À3 (t ¼ r= ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ffi ð1 À r 2 Þ=ðN À 2Þ p , t-test). Throughout testing, trends remained when both the length of the x-vector and the correlation lower limit were either strengthened or weakened.

Global convolution-supervised clustering
The goal of our clustering approach was to find transcripts that would be biologically easy to explain. In particular, we were interested in tying clusters of transcripts to specific stages of development, such as germ-band elongation or neurogenesis. Our strategy was therefore to cluster transcripts that had very consistent expression patterns. Consistency in this case would mean an unbroken state of, for example, high expression followed by an unbroken state of low expression, such as for the maternally inherited transcripts. The actual fold change of the transcript levels was not as important as the consistency requirement, meaning that we could study genes with less dramatic transcript levels than traditional clustering. The disadvantages of traditional clustering include (a) random transcript spikes leading to predictions that were difficult to assess biologically and (b) effects of noisy data points leading to false positives. Using a supervised clustering technique would classify these cases as true negatives.
When searching for plateaus of expression, we convolved the expression profiles according to S i,j ¼c(e, x i,j ), where e is the expression profile vector and Here, c is the correlation coefficient function. The expression profile is considered to be active from i, y, j if the maximum of S exceeds 0.8 (Po10 À4 , t-test). For example, Supplementary Figure S1 illustrates a high correlation between an expression profile and the filter vector x i,j , where i¼18 and j¼25. The maximum value of S in this case is 0.95. This approach is conceptually similar to the method employed by Šášik et al (2002), although here we actively look for steady plateaus followed by sharp declines in expression. Furthermore, an advantage of using correlation coefficients is that P-values are given, removing the need for random sampling. This method of clustering was chosen as it does not assume Euclidean distance between genes. A Euclidean distance implies that there is no time dependency, which is not consistent with our expectations. Finally, as mentioned above, the resulting clusters are more readily explained biologically as contiguous phases of up-and downregulation. For class II, in order to distinguish transcript expression profiles from those that are maternally deposited, we set the first six data points to low expression and varied the remaining 24. Hence, the number of combinations is 276 and not 435. For class III, we required that expression be still high at the end of the time series.

In situ data
The in situ data were retrieved from the Berkeley database (Tomancak et al, 2002). This database contains in situ data annotation for 2152 genes with 211 anatomical terms that are based on pictures taken at five different developmental times. Data are available for 253 out of our 842 significantly correlated genes.

Pathways
A total of 1309 genes grouped in 31 pathways were retrieved from the Interactive Fly database (http://flybase.bio.indiana.edu/allied-data/ lk/interactive-fly/aimain/1aahome.htm). Out of these, 86% are spotted on the microarray.

Orthology
Orthologs in A. gambiae were retrieved from the STRING database (von Mering et al, 2005).

Lethality
We consider all genes as lethal that were annotated as 'Phenotypic class: lethal' in Flybase (Drysdale and Crosby, 2005) unless the time of manifestation was given and stated manifestation of the phenotype only in the larvae or later stages. In total, 711 genes of the 12 868 genes (5.5%) tested were considered as lethal.

Transcription factors
A total of 336 genes annotated as 'transcription factor' were extracted from FlyBase (Drysdale and Crosby, 2005) and manually curated (Tobias Doerks, personal communication).

GO annotation
Overrepresentation of GO categories was analyzed using the GOSSIP program (Bluthgen et al, 2005), correcting for multiple testing using FDR and FWER.

Protein-protein interactions
The interaction network was obtained from release 6.2 of the STRING database (von Mering et al, 2005), which includes curated data from the yeast two-hybrid screen by Giot et al (2003) and many individually reported interactions. All interaction data for D. melanogaster were used with the exception of links inferred from mRNA coexpression data (Arbeitman et al, 2002). The cutoff for STRING scores was set at 0.3. The network figure was created using Cytoscape 2.1 (Shannon et al, 2003) and Medusa .

PEST degradation signals
The PEST-find program (Rechsteiner and Rogers, 1996) was used to perform proteome-wide computational search for PEST regions. The number of PEST-containing proteins was counted among all significantly regulated genes as well as within each 'expression class'. The statistical significance of PEST overrepresentation was calculated for each 'expression class' compared to all significantly regulated genes using the exact hypergeometric test.

Supplementary information
Supplementary information is available at the Molecular Systems Biology website (www.nature.com/msb).